Class-wise Knowledge Distillation for Lightweight Segmentation Model
Class-wise Knowledge Distillation for Lightweight Segmentation Model
复制标题
DOI：
10.5220/0011719900003414
复制
发表时间：
2023
期刊：
影响因子：
--
通讯作者：
Ryota Ikedo;Kotaro Nagata;K. Hotta
中科院分区：
文献类型：
--
作者：
Ryota Ikedo;Kotaro Nagata;K. Hotta
近年来，我们一直在通过深化切分模型来提高语义切分的准确性，但由于计算复杂度的增加，需要大量的计算资源。因此，知识提取作为一种模型压缩方法得到了研究。我们提出了一种知识提取方法，将每个班级学习到的教师模型的输出分布作为学生模型的目标，以达到压缩记忆和提高准确率的目的。实验结果表明，在不增加两个不同数据集的计算代价的情况下，提高了分割精度。
: In recent years, we have been improving the accuracy of semantic segmentation by deepening segmentation models, but large amount of computational resources are required due to the increase in computational complexity. Therefore knowledge distillation has been studied as one of model compression methods. We propose a knowledge distillation method in which the output distribution of a teacher model learned for each class is used as a target of the student model for the purpose of memory compression and accuracy improvement. Experimental results demonstrate that the segmentation accuracy was improved without increasing the computational cost on two different datasets.